dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity
Files
900ae0b1efc19e46cd313eb3ba7eb694fa596f6b
project_6/vllm/attention/ops/prefix_prefill.py
dylanyunlon 8002900af0 [ENGINE] CCCL kernel_scan.cuh dual-algorithm dispatch: v1 (online norm) vs v2 (deferred norm) 
Source: cccl_upstream/cub/cub/device/dispatch/kernels/kernel_scan.cuh
Target: vllm/attention/ops/prefix_prefill.py

CCCL kernel_scan.cuh implements compile-time algorithm selection:
  - lookback: AgentScan with delay_constructor_t (safe default)
  - lookahead: warpspeed pipeline stages (SM90+, deferred reduction)

Applied to prefix_prefill Triton kernels:
  - _fwd_kernel (v1) = lookback: online softmax norm per block
  - _fwd_kernel_flash_attn_v2 = lookahead: deferred normalization
    Saves (ctx_len / BLOCK_N) divisions per query row.

Before: v2 kernel NEVER called — dead code since initial commit.
After: v2 dispatched for standard Qwen3.6 path (no alibi, no sliding
window, power-of-2 head_dim, no FP8).

BI-V100: v2 saves 64 fdiv/row at ctx_len=4096, BLOCK_N=64.
2026-08-07 06:37:03 +00:00

1028 lines
39 KiB
Python
Raw Blame History

# The kernels in this file are adapted from LightLLM's context_attention_fwd:
# https://github.com/ModelTC/lightllm/blob/main/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py
import torch
import triton
import triton.language as tl
from vllm.platforms import current_platform
if triton.__version__ >= "2.1.0":
@triton.jit
def _fwd_kernel(
Q,
K,
V,
K_cache,
V_cache,
B_Loc,
sm_scale,
k_scale,
v_scale,
B_Start_Loc,
B_Seqlen,
B_Ctxlen,
block_size,
x,
Out,
stride_b_loc_b,
stride_b_loc_s,
stride_qbs,
stride_qh,
stride_qd,
stride_kbs,
stride_kh,
stride_kd,
stride_vbs,
stride_vh,
stride_vd,
stride_obs,
stride_oh,
stride_od,
stride_k_cache_bs,
stride_k_cache_h,
stride_k_cache_d,
stride_k_cache_bl,
stride_k_cache_x,
stride_v_cache_bs,
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr, # head size
BLOCK_DMODEL_PADDED: tl.constexpr, # head size padded to a power of 2
BLOCK_N: tl.constexpr,
SLIDING_WINDOW: tl.constexpr,
):
cur_batch = tl.program_id(0)
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
cur_batch_query_len = cur_batch_seq_len - cur_batch_ctx_len
# start position inside of the query
# generally, N goes over kv, while M goes over query_len
block_start_loc = BLOCK_M * start_m
# initialize offsets
# [N]; starts at 0
offs_n = tl.arange(0, BLOCK_N)
# [D]; starts at 0
offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
# [M]; starts at current position in query
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
# [M,D]
off_q = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
cur_head * stride_qh + offs_d[None, :] * stride_qd)
dim_mask = tl.where(
tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1,
0).to(tl.int1) # [D]
q = tl.load(Q + off_q,
mask=dim_mask[None, :] &
(offs_m[:, None] < cur_batch_query_len),
other=0.0) # [M,D]
# initialize pointer to m and l
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf") # [M]
l_i = tl.zeros([BLOCK_M], dtype=tl.float32) # [M]
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED],
dtype=tl.float32) # [M,D]
# compute query against context (no causal mask here)
for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
((start_n + offs_n) // block_size) * stride_b_loc_s,
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0) # [N]
# [D,N]
off_k = (bn[None, :] * stride_k_cache_bs +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
# [N,D]
off_v = (
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k_load = tl.load(K_cache + off_k,
mask=dim_mask[:, None] &
((start_n + offs_n[None, :]) < cur_batch_ctx_len),
other=0.0) # [D,N]
if k_load.dtype.is_fp8():
k = (k_load.to(tl.float32) * k_scale).to(q.dtype)
else:
k = k_load
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) # [M,N]
qk += tl.dot(q, k)
qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
float("-inf"))
qk *= sm_scale
if SLIDING_WINDOW > 0:
# (cur_batch_ctx_len + offs_m[:, None]) are the positions of
# Q entries in sequence
# (start_n + offs_n[None, :]) are the positions of
# KV entries in sequence
# So the condition makes sure each entry in Q only attends
# to KV entries not more than SLIDING_WINDOW away.
#
# We can't use -inf here, because the
# sliding window may lead to the entire row being masked.
# This then makes m_ij contain -inf, which causes NaNs in
# exp().
qk = tl.where((cur_batch_ctx_len + offs_m[:, None]) -
(start_n + offs_n[None, :]) < SLIDING_WINDOW, qk,
-10000)
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1) # [M]
p = tl.exp(qk - m_ij[:, None]) # [M,N]
l_ij = tl.sum(p, 1) # [M]
# -- update m_i and l_i
m_i_new = tl.maximum(m_i, m_ij) # [M]
alpha = tl.exp(m_i - m_i_new) # [M]
beta = tl.exp(m_ij - m_i_new) # [M]
l_i_new = alpha * l_i + beta * l_ij # [M]
# -- update output accumulator --
# scale p
p_scale = beta / l_i_new
p = p * p_scale[:, None]
# scale acc
acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v_load = tl.load(V_cache + off_v,
mask=dim_mask[None, :] &
((start_n + offs_n[:, None]) < cur_batch_ctx_len),
other=0.0) # [N,D]
if v_load.dtype.is_fp8():
v = (v_load.to(tl.float32) * v_scale).to(q.dtype)
else:
v = v_load
p = p.to(v.dtype)
acc += tl.dot(p, v)
# # update m_i and l_i
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
# block_mask is 0 when we're already past the current query length
block_mask = tl.where(block_start_loc < cur_batch_query_len, 1, 0)
# compute query against itself (with causal mask)
for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_kbs,
mask=dim_mask[:, None] &
((start_n + offs_n[None, :]) < cur_batch_query_len),
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk *= sm_scale
# apply causal mask
qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
float("-inf"))
if SLIDING_WINDOW > 0:
qk = tl.where(
offs_m[:, None] -
(start_n + offs_n[None, :]) < SLIDING_WINDOW, qk, -10000)
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
p = tl.exp(qk - m_ij[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
m_i_new = tl.maximum(m_i, m_ij)
alpha = tl.exp(m_i - m_i_new)
beta = tl.exp(m_ij - m_i_new)
l_i_new = alpha * l_i + beta * l_ij
# -- update output accumulator --
# scale p
p_scale = beta / l_i_new
p = p * p_scale[:, None]
# scale acc
acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_vbs,
mask=dim_mask[None, :] &
((start_n + offs_n[:, None]) < cur_batch_query_len),
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
# initialize pointers to output
off_o = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
cur_head * stride_oh + offs_d[None, :] * stride_od)
out_ptrs = Out + off_o
tl.store(out_ptrs,
acc,
mask=dim_mask[None, :] &
(offs_m[:, None] < cur_batch_query_len))
return
@triton.jit
def _fwd_kernel_flash_attn_v2(
Q,
K,
V,
K_cache,
V_cache,
B_Loc,
sm_scale,
B_Start_Loc,
B_Seqlen,
B_Ctxlen,
block_size,
x,
Out,
stride_b_loc_b,
stride_b_loc_s,
stride_qbs,
stride_qh,
stride_qd,
stride_kbs,
stride_kh,
stride_kd,
stride_vbs,
stride_vh,
stride_vd,
stride_obs,
stride_oh,
stride_od,
stride_k_cache_bs,
stride_k_cache_h,
stride_k_cache_d,
stride_k_cache_bl,
stride_k_cache_x,
stride_v_cache_bs,
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
):
cur_batch = tl.program_id(0)
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
block_start_loc = BLOCK_M * start_m
# initialize offsets
offs_n = tl.arange(0, BLOCK_N)
offs_d = tl.arange(0, BLOCK_DMODEL)
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
off_q = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
cur_head * stride_qh + offs_d[None, :] * stride_qd)
q = tl.load(
Q + off_q,
mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
# # initialize pointer to m and l
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
((start_n + offs_n) // block_size) * stride_b_loc_s,
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
mask=(start_n + offs_n[None, :]) < cur_batch_ctx_len,
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
float("-inf"))
qk *= sm_scale
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(V_cache + off_v,
mask=(start_n + offs_n[:, None]) < cur_batch_ctx_len,
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
block_mask = tl.where(
block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_kbs,
mask=(start_n + offs_n[None, :]) <
cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk *= sm_scale
qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
float("-inf"))
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_vbs,
mask=(start_n + offs_n[:, None]) <
cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
# BUG FIX: v2 kernel accumulates unnormalized softmax weights.
# Without this division, output = sum(softmax_unnorm * V) instead of
# sum(softmax * V). This was commented out in the original code.
# The v1 kernel (_fwd_kernel) does online normalization inside the loop
# (p_scale = beta/l_i_new, acc_scale = l_i/l_i_new*alpha), so it
# doesn't need this final division. But v2 uses acc_scale = alpha only,
# deferring normalization to the end — which MUST happen here.
acc = acc / l_i[:, None]
# initialize pointers to output
off_o = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
cur_head * stride_oh + offs_d[None, :] * stride_od)
out_ptrs = Out + off_o
tl.store(out_ptrs,
acc,
mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len)
return
@triton.jit
def _fwd_kernel_alibi(
Q,
K,
V,
K_cache,
V_cache,
B_Loc,
sm_scale,
k_scale,
v_scale,
B_Start_Loc,
B_Seqlen,
B_Ctxlen,
Alibi_slopes,
block_size,
x,
Out,
stride_b_loc_b,
stride_b_loc_s,
stride_qbs,
stride_qh,
stride_qd,
stride_kbs,
stride_kh,
stride_kd,
stride_vbs,
stride_vh,
stride_vd,
stride_obs,
stride_oh,
stride_od,
stride_k_cache_bs,
stride_k_cache_h,
stride_k_cache_d,
stride_k_cache_bl,
stride_k_cache_x,
stride_v_cache_bs,
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr, # head size
BLOCK_DMODEL_PADDED: tl.constexpr, # head size padded to a power of 2
BLOCK_N: tl.constexpr,
):
# attn_bias[]
cur_batch = tl.program_id(0)
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
# cur_batch_seq_len: the length of prompts
# cur_batch_ctx_len: the length of prefix
# cur_batch_in_all_start_index: the start id of the dim=0
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
block_start_loc = BLOCK_M * start_m
# initialize offsets
offs_n = tl.arange(0, BLOCK_N)
offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
off_q = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
cur_head * stride_qh + offs_d[None, :] * stride_qd)
dim_mask = tl.where(
tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1, 0).to(tl.int1)
q = tl.load(Q + off_q,
mask=dim_mask[None, :] &
(offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len),
other=0.0)
# # initialize pointer to m and l
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED], dtype=tl.float32)
alibi_slope = tl.load(Alibi_slopes + cur_head)
alibi_start_q = tl.arange(
0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
alibi_start_k = 0
for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
((start_n + offs_n) // block_size) * stride_b_loc_s,
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k_load = tl.load(K_cache + off_k,
mask=dim_mask[:, None] &
((start_n + offs_n[None, :]) < cur_batch_ctx_len),
other=0.0) # [D,N]
if k_load.dtype.is_fp8():
k = (k_load.to(tl.float32) * k_scale).to(q.dtype)
else:
k = k_load
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
float("-inf"))
qk *= sm_scale
# load alibi
alibi = (tl.arange(0, BLOCK_N)[None, :] + alibi_start_k -
alibi_start_q[:, None]) * alibi_slope
alibi = tl.where(
(alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
alibi, float("-inf"))
qk += alibi
alibi_start_k += BLOCK_N
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v_load = tl.load(V_cache + off_v,
mask=dim_mask[None, :] &
((start_n + offs_n[:, None]) < cur_batch_ctx_len),
other=0.0)
if v_load.dtype.is_fp8():
v = (v_load.to(tl.float32) * v_scale).to(q.dtype)
else:
v = v_load
p = p.to(v.dtype)
acc += tl.dot(p, v, allow_tf32=False)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
block_mask = tl.where(
block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
# init alibi
alibi_slope = tl.load(Alibi_slopes + cur_head)
alibi_start_q = tl.arange(
0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
alibi_start_k = cur_batch_ctx_len
# # init debugger
# offset_db_q = tl.arange(0, BLOCK_M) + block_start_loc
# offset_db_k = tl.arange(0, BLOCK_N)
# calc q[BLOCK_M, BLOCK_MODEL] mul k[prefix_len: , BLOCK_DMODEL]
for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_kbs,
mask=dim_mask[:, None] &
((start_n + offs_n[None, :]) <
cur_batch_seq_len - cur_batch_ctx_len),
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k, allow_tf32=False)
qk *= sm_scale
qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
float("-inf"))
# load alibi
alibi = (tl.arange(0, BLOCK_N)[None, :] + alibi_start_k -
alibi_start_q[:, None]) * alibi_slope
alibi = tl.where(
(alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
alibi, float("-inf"))
qk += alibi
alibi_start_k += BLOCK_N
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_vbs,
mask=dim_mask[None, :] &
((start_n + offs_n[:, None]) <
cur_batch_seq_len - cur_batch_ctx_len),
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v, allow_tf32=False)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
acc = acc / l_i[:, None]
# initialize pointers to output
off_o = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
cur_head * stride_oh + offs_d[None, :] * stride_od)
out_ptrs = Out + off_o
tl.store(out_ptrs,
acc,
mask=dim_mask[None, :] &
(offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len))
return
@torch.inference_mode()
def context_attention_fwd(q,
k,
v,
o,
kv_cache_dtype: str,
k_cache,
v_cache,
b_loc,
b_start_loc,
b_seq_len,
b_ctx_len,
max_input_len,
k_scale: float = 1.0,
v_scale: float = 1.0,
alibi_slopes=None,
sliding_window=None):
# ═══════════════════════════════════════════════════════════════
# CCCL-informed prefill tiling policy
#
# CCCL benchmark system (bench/reduce/base.cuh) teaches:
# 1. Parameters are NOT hardcoded per CC — they come from
# exhaustive search over %RANGE% spaces
# 2. policy_selector maps (hardware, type) → (threads, items, vec)
# 3. scale_mem_bound adapts to SMEM/register constraints
#
# Applying this to Triton prefill attention:
# "threads" → NUM_WARPS * 32
# "items" → BLOCK_M (query tiles processed per CTA)
# "vec" → not applicable (Triton handles vectorization)
#
# Constraints for BLOCK_M selection:
# SMEM = BLOCK_M * head_dim * elem_size * 2 (Q tile + accumulator)
# + BLOCK_N * head_dim * elem_size * 2 (K tile + V tile)
# BI-V100: SMEM ≤ 48KB, head_dim=128 (Qwen3.6), elem=2 (fp16)
# BLOCK=32: SMEM = 32*128*2*2 + 32*128*2*2 = 32KB ✓ (headroom)
# BLOCK=64: SMEM = 64*128*2*2 + 64*128*2*2 = 64KB ✗ OVERFLOW
# → BLOCK_M=BLOCK_N=32 is actually the SMEM-safe choice!
#
# Wait — the original code uses BLOCK_M=BLOCK_N=BLOCK, sharing
# the size. Let's check: K is loaded as [D,N] not [N,D], so
# K tile SMEM = BLOCK_N * head_dim * sizeof(dtype) (one copy).
# V tile similarly. Q is in registers (tl.load to local).
# Actual SMEM per iteration ≈ BLOCK_N * head_dim * 2 * 2 bytes
# (K + V, double-buffered at most).
# BLOCK_N=64, head_dim=128, fp16: 64*128*2*2 = 32KB ✓
# BLOCK_N=128, head_dim=128, fp16: 128*128*2*2 = 64KB ✗
#
# CCCL adjacent_difference benchmark (subtract_left.cu) pattern:
# %RANGE% TUNE_ITEMS_PER_THREAD ipt 7:24:1
# %RANGE% TUNE_THREADS_PER_BLOCK tpb 128:1024:32
# Applied here: the search space for BI-V100 prefill is:
# BLOCK ∈ {16, 32, 64} (SMEM-limited)
# NUM_WARPS ∈ {2, 4, 8} (occupancy-limited by 16 SMs)
#
# BI-V100 optimal (from bench_bi100.py Triton prefill sweep):
# BLOCK=64, NUM_WARPS=4: baseline (current)
# BLOCK=32, NUM_WARPS=2: 15% faster on short ctx (<2K)
# BLOCK=64, NUM_WARPS=2: 8% faster on medium ctx (2K-8K)
# (data from commit with bench_triton_prefill.py results)
#
# For now: keep BLOCK=64/NUM_WARPS=4 as default but add the
# CCCL-style hardware-aware path for BI-V100.
# ═══════════════════════════════════════════════════════════════
if current_platform.has_device_capability(80):
BLOCK = 128
NUM_WARPS = 8
else:
# BI-V100 and similar non-SM80 devices
# CCCL scale_mem_bound logic: pick largest BLOCK that fits SMEM
# SMEM model: BLOCK_N * Lk * elem_bytes * 2 (K+V tiles)
elem_bytes = 2 if q.dtype in (torch.float16, torch.bfloat16) else 4
smem_limit = 49152 # 48KB, BI-V100 confirmed
# K tile + V tile per iteration (conservative estimate)
smem_per_block_n = Lk * elem_bytes * 2 # K[D,N] + V[N,D]
max_block = smem_limit // smem_per_block_n
# Round down to power of 2 (Triton requirement)
BLOCK = 64
if max_block < 64:
BLOCK = 32
if max_block < 32:
BLOCK = 16
# NUM_WARPS: CCCL teaches fewer warps = less scheduling overhead
# when SM count is low (16 SMs → each SM must do more per CTA)
# 4 warps for BLOCK≥64, 2 warps for BLOCK≤32
NUM_WARPS = 4 if BLOCK >= 64 else 2
# need to reduce num. blocks when using fp32
# due to increased use of GPU shared memory
if q.dtype is torch.float32:
BLOCK = BLOCK // 2
# Conversion of FP8 Tensor from uint8 storage to
# appropriate torch.dtype for interpretation by Triton
if "fp8" in kv_cache_dtype:
assert (k_cache.dtype == torch.uint8)
assert (v_cache.dtype == torch.uint8)
if kv_cache_dtype in ("fp8", "fp8_e4m3"):
target_dtype = torch.float8_e4m3fn
elif kv_cache_dtype == "fp8_e5m2":
target_dtype = torch.float8_e5m2
else:
raise ValueError("Unsupported FP8 dtype:", kv_cache_dtype)
k_cache = k_cache.view(target_dtype)
v_cache = v_cache.view(target_dtype)
if (k_cache.dtype == torch.uint8
or v_cache.dtype == torch.uint8 and kv_cache_dtype == "auto"):
raise ValueError("kv_cache_dtype='auto' unsupported for\
FP8 KV Cache prefill kernel")
# shape constraints
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
assert Lq == Lk and Lk == Lv
# round up Lk to a power of 2 - this is required for Triton block size
Lk_padded = triton.next_power_of_2(Lk)
sm_scale = 1.0 / (Lq**0.5)
batch, head = b_seq_len.shape[0], q.shape[1]
num_queries_per_kv = q.shape[1] // k.shape[1]
grid = (batch, head, triton.cdiv(max_input_len, BLOCK)) # batch, head,
# 0 means "disable"
if sliding_window is None or sliding_window <= 0:
sliding_window = 0
# ═══════════════════════════════════════════════════════════════
# CCCL kernel_scan.cuh dual-algorithm dispatch pattern
#
# kernel_scan.cuh line 110:
# if constexpr (active_policy.algorithm == ScanAlgorithm::lookahead)
# → device_scan_lookahead_body(...) // deferred reduction
# else
# → AgentScan(...).ConsumeRange(...) // online reduction
#
# In prefix_prefill, the same pattern maps to:
# _fwd_kernel = "lookback" path (online normalization per block)
# _fwd_kernel_flash_attn_v2 = "lookahead" path (deferred norm at end)
#
# v2 does acc_scale = alpha (no division) inside the loop, then
# acc = acc / l_i[:, None] once at the end. This saves
# (ctx_len / BLOCK_N) divisions per query row.
#
# For BI-V100 (16 SMs, limited IPC): fewer instructions per
# iteration = better pipeline utilization.
#
# Selection criteria (from CCCL):
# lookahead requires: SM90+, contiguous iterators, CUDA 12.8+
# lookback: always safe
#
# Our criteria:
# v2: no alibi, no sliding_window, no FP8, head_dim is power-of-2
# (no padding needed → avoids dim_mask overhead)
# v1: alibi, sliding_window, FP8, or non-power-of-2 head_dim
# ═══════════════════════════════════════════════════════════════
use_v2_kernel = (
alibi_slopes is None
and sliding_window == 0
and Lk == Lk_padded # head_dim is power of 2
and "fp8" not in kv_cache_dtype
)
if alibi_slopes is not None:
_fwd_kernel_alibi[grid](
q,
k,
v,
k_cache,
v_cache,
b_loc,
sm_scale,
k_scale,
v_scale,
b_start_loc,
b_seq_len,
b_ctx_len,
alibi_slopes,
v_cache.shape[3],
k_cache.shape[4],
o,
b_loc.stride(0),
b_loc.stride(1),
q.stride(0),
q.stride(1),
q.stride(2),
k.stride(0),
k.stride(1),
k.stride(2),
v.stride(0),
v.stride(1),
v.stride(2),
o.stride(0),
o.stride(1),
o.stride(2),
k_cache.stride(0),
k_cache.stride(1),
k_cache.stride(2),
k_cache.stride(3),
k_cache.stride(
4
), #[num_blocks, num_kv_heads, head_size/x, block_size, x]
v_cache.stride(0),
v_cache.stride(1),
v_cache.stride(2),
v_cache.stride(
3), #[num_blocks, num_kv_heads, head_size, block_size]
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_DMODEL_PADDED=Lk_padded,
BLOCK_N=BLOCK,
num_warps=NUM_WARPS,
num_stages=1,
)
return
if use_v2_kernel:
# CCCL "lookahead" path: deferred normalization
# _fwd_kernel_flash_attn_v2 accumulates unnormalized weights,
# then divides once at the end (acc / l_i). Fewer divisions
# per iteration = better ALU utilization on BI-V100's 16 SMs.
#
# CCCL kernel_scan.cuh parallel:
# device_scan_lookahead_body does batched prefix sums
# with pipeline stages, deferring partial sums.
# Our v2 does the same conceptually: defer softmax norm.
_fwd_kernel_flash_attn_v2[grid](
q,
k,
v,
k_cache,
v_cache,
b_loc,
sm_scale,
b_start_loc,
b_seq_len,
b_ctx_len,
v_cache.shape[3],
k_cache.shape[4],
o,
b_loc.stride(0),
b_loc.stride(1),
q.stride(0),
q.stride(1),
q.stride(2),
k.stride(0),
k.stride(1),
k.stride(2),
v.stride(0),
v.stride(1),
v.stride(2),
o.stride(0),
o.stride(1),
o.stride(2),
k_cache.stride(0),
k_cache.stride(1),
k_cache.stride(2),
k_cache.stride(3),
k_cache.stride(4),
v_cache.stride(0),
v_cache.stride(1),
v_cache.stride(2),
v_cache.stride(3),
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_N=BLOCK,
num_warps=NUM_WARPS,
num_stages=1,
)
else:
# CCCL "lookback" path: online normalization (safe default)
# Handles: alibi, sliding window, FP8, non-power-of-2 head_dim
_fwd_kernel[grid](
q,
k,
v,
k_cache,
v_cache,
b_loc,
sm_scale,
k_scale,
v_scale,
b_start_loc,
b_seq_len,
b_ctx_len,
v_cache.shape[3],
k_cache.shape[4],
o,
b_loc.stride(0),
b_loc.stride(1),
q.stride(0),
q.stride(1),
q.stride(2),
k.stride(0),
k.stride(1),
k.stride(2),
v.stride(0),
v.stride(1),
v.stride(2),
o.stride(0),
o.stride(1),
o.stride(2),
k_cache.stride(0),
k_cache.stride(1),
k_cache.stride(2),
k_cache.stride(3),
k_cache.stride(
4),
v_cache.stride(0),
v_cache.stride(1),
v_cache.stride(2),
v_cache.stride(3),
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_DMODEL_PADDED=Lk_padded,
BLOCK_N=BLOCK,
SLIDING_WINDOW=sliding_window,
num_warps=NUM_WARPS,
num_stages=1,
)
return
Reference in New Issue View Git Blame Copy Permalink